Deconvolving RNA Base Pairing Signals

The structure of an RNA sequence encodes information about its biological function. A sequence is typically predicted to fold to a single minimum free energy conformation. But, an increasing number of RNA molecules are now known to fold into multiple stable structures. Discrete optimization methods are commonly used to predict foldings, and adding experimental data as auxiliary information improves prediction accuracy when there is a single dominant conformation. In this paper, we analyze the outputs of existing structural prediction models when they receive auxiliary data derived from a mixture of structures. Under a binary model of auxiliary data, we find that current structural prediction methods typically favor distributions with one dominant structure, and hence cannot guarantee accurate reconstruction of multimodal distributions. Additionally, we analyze empirical distributions of auxiliary data used in current prediction models. We show that even when the structures in a distribution are known in advance, it is difficult to determine the weightings of the structures using auxiliary data. RNA secondary structure and thermodynamic optimization with auxiliary data and method of moments estimators